AI content and search: the real risk for a small site

The question we get asked most is whether Google punishes a page for being written by AI. It does not, and it has never said it does. What the spam policies page says, in the version last updated on 28 August 2026, is this: “Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users.” Generative AI appears there once, as an example of a way to do that, alongside older methods.

What carries the risk is the behaviour the tool makes cheap.

What Google has actually been doing about it

Three spam updates ran between late June and late September. The Search status history dates them to 24 June, 18 August and 24 September 2026, on top of core updates in March (twelve days) and May (nearly twelve days). That is a faster cadence than the previous two years, and if you publish at volume you are now rolling dice roughly every eight weeks.

The pattern that gets caught is recognisable without any special tooling:

• Forty pages in a fortnight on near-identical topics
• Location pages where the only variable is the city name
• An article that answers the query by restating the current first page of results
• No author, no date, no specifics, nothing that could only have come from this business
• Published unedited, which is usually obvious by the second paragraph

The things being sold to you that Google says do nothing

This is the useful part, because an entire consulting category has grown up around “AEO” and “GEO” in the past eighteen months and Google has now written down which bits of it are fiction. Its guide to optimising for generative AI features, last updated 10 July 2026, is blunt.

An llms.txt file does nothing: Google says Search does not use them. Breaking your content into “chunks” for retrieval is unnecessary. Writing in a special register for machines is unnecessary too, in Google’s words, “You don’t need to write in a specific way just for generative AI search.” Chasing brand mentions you have not earned does not work. The separate AI features documentation puts it plainly: there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimisations necessary.

If someone has quoted you a monthly fee to install an llms.txt and restructure your pages into machine-readable chunks, that is the invoice to query.

What the same guide does ask for is content that is not a commodity. Its own example contrasts “7 Tips for First-Time Homebuyers”, which anyone could write, with “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line”, which only someone who did it could. The instruction is to avoid recycling what is already on the internet or what a model could produce unaided.

Now the numbers, and what is wrong with them

We held these back deliberately, because they get quoted far more confidently than they deserve.

The best independent measurement of behaviour is Pew Research’s study of real browsing data from 900 US adults, covering 68,879 Google searches, of which 12,593 produced an AI summary. Users clicked a result 8% of the time when a summary was present and 15% when it was not. They clicked a link inside the summary 1% of the time. They ended the browsing session entirely on 26% of visits with a summary, against 16% without.

That data was collected in March 2025. It is eighteen months old, it is US-only, and AI Mode did not exist in its current form when it was gathered.

Ahrefs published an update in February 2026 putting the drop in clickthrough for the top-ranking page at 58% where an AI Overview is present, up from the 34.5% they measured in 2025. Their method compares aggregated Search Console clickthrough across 300,000 keywords, December 2023 against December 2025. The authors state clearly that this is a correlation and that two years of other changes sit inside the comparison. We think it is the most careful large-scale attempt available and still would not quote it as your expected loss.

For what it is worth, Google’s own position is that clicks arriving from pages with AI Overviews are “higher quality”, meaning visitors stay longer. It has not published the data behind that.

The question we cannot answer

Here is the one we keep running into and have no good answer for. Being cited in an AI answer is now the thing everyone is told to optimise for. Nobody has shown us that a citation is worth money.

Pew found a 1% click rate on cited sources. If a citation mostly produces an unclicked mention, its value has to come from something else, brand recall, or a visitor who searches your name later, and we have not seen that measured for a business of five or fifteen people. Cloudflare’s published figures, covering January to July 2025, show how lopsided the exchange has become on the crawl side: Anthropic’s crawler fetched roughly 38,000 pages per referral it sent back by July, OpenAI’s about 1,091, Google’s about 5. We have watched those numbers get relabelled as 2026 data in a dozen blog posts. They are 2025 data.

So we do not know. Anyone who tells you they have a reliable model of what an AI citation is worth to a small business is ahead of the evidence, and we would rather leave the question standing than fill it with a figure.

What we do in the meantime

Nothing exotic, and most of it predates all of this. Topics come from questions real customers asked, from sales calls and from the search terms in your own Search Console, never from a model’s suggestion of what to write. The specifics go in before the drafting: your prices, your process, the objection you hear every week, the job that went wrong. If a paragraph would be true for any competitor in your category, it comes out. The review time this costs is the same trap we described in where AI earns its cost and where it quietly does not, and content is where we see it most often.

Business details stay identical across the site, the Google Business Profile and every directory, because inconsistency makes you an unreliable thing to quote. Every page carries a named author and a visible date. And we publish at a human rate, which for most of our clients means one page a week or less. This is the slow end of what we do in a consultation, and it is usually a shorter list than people expect.

The real cost of getting this wrong is a wasted quarter: thirty pages that earn nothing, followed by the conclusion that content does not work for your business. The pages stay up afterwards with your name on them, readable by anyone deciding whether to hire you.